/consensus-validate

Activate the consensus-pipeline skill for a single-phase gate check. Unlike /consensus-run, this produces evidence but does NOT mutate $target/.consensus/state.json — useful for spot-checks after a targeted fix. Ports src/consensus/cli.py:101-147 (validate subcommand) semantics.

Arguments

Parse $ARGUMENTS:

• --phase PHASE (required) — the phase to validate (e.g., audit)
• --target PATH — path to target repo (default: .)

Execution

1. Read config from $TARGET/.claude/consensus.local.md (fall back to defaults from src/consensus/config.py:29-56).
2. Verify PHASE is a known phase name per config + defaults. Exit non-zero if unknown.
3. Run bash ${CLAUDE_PLUGIN_ROOT}/skills/consensus-pipeline/scripts/init-evidence-dir.sh $TARGET $PHASE.
4. Spawn Lead, Alpha, Bravo in one parallel tool-use block via the Task tool (same independence rules as /consensus-run).
5. Run bash ${CLAUDE_PLUGIN_ROOT}/skills/consensus-pipeline/scripts/gate-check.sh $TARGET $PHASE 1 (single gate, gate_number=1).
6. Print the verdict from $TARGET/.consensus/evidence/$PHASE/gate-1-verdict.json. Do NOT create or modify $TARGET/.consensus/state.json — this command is a spot-check, not a pipeline run.

Exit 0 on unanimous PASS, 1 on any FAIL, 2 on hard error (malformed JSON / missing votes).

Pre-Flight Canary — cheap pre-check before full /consensus-run

A full /consensus-run over the default 4-phase pipeline (explore,audit,fix,verify) with a fix cycle spends roughly 24–30 LLM completions — 3 agents × 4 phases × up to 2 gates per phase. That is the expensive path.

Use /consensus-validate --phase audit FIRST when you want to spot-check agent calibration before committing to the full burn. The canary:

• spawns exactly 3 agents (Lead + Alpha + Bravo) once,
• runs one gate,
• costs ~3 LLM completions,
• produces $TARGET/.consensus/evidence/audit/gate-1-*.json you can inspect before deciding.

Recommended flow for new targets:

1. /consensus-validate --phase audit --target <TARGET> (canary, ~3 completions).
2. Inspect the alpha vote for the defect you expect — if Alpha does not FAIL citing the known issue, your agents are miscalibrated or the sample is wrong. Fix before running the full pipeline.
3. On a clean canary PASS (or an expected FAIL with the right citation), run /consensus-run --target <TARGET> for the full pipeline.

This pattern is load-bearing for budget protection: it catches sample-validity and prompt-interpolation bugs BEFORE the ~30-completion phase-04 dogfood run.